fix: slurm host system access

cobald.yml

@@ -43,6 +43,7 @@
         # containers runtime name (not hostname supplied!) and netname is
         # the network name in host environment. We should run our own dns...
       docker_network: slurm
+      slurm_hostsystem_cluster_access: True
     when: '"slurm" in group_names'
     tags: slurm, cobald, influxdb, slurm-config
     # tags: cobald requires some slurm facts, so cobald tag is included here

roles/cobald/files/cgroup.conf.noautomount

@@ -0,0 +1,11 @@
+###
+#
+# Slurm cgroup support configuration file
+#
+# See man slurm.conf and man cgroup.conf for further
+# information on cgroup configuration parameters
+#--
+CgroupAutomount=no
+
+ConstrainCores=no
+ConstrainRAMSpace=no

roles/cobald/files/cobald-config/config.yaml

@@ -0,0 +1,31 @@
+---
+pipeline:
+- __type__: cobald.controller.linear.LinearController
+  low_utilisation: 0.9
+  high_allocation: 0.9
+  rate: 0.10
+- !Limiter
+  minimum: 3
+- !TelegrafPipelineMonitor
+  poll: True
+- !TardisPoolFactory
+  configuration: /etc/cobald/tardis.yaml
+logging:
+  version: 1
+  root:
+    level: DEBUG
+    handlers: [console, file]
+  handlers:
+    console:
+      class: logging.StreamHandler
+      formatter: test
+      level: DEBUG
+      stream: ext://sys.stderr
+    file:
+      class: logging.handlers.RotatingFileHandler
+      formatter: test
+      level: WARNING
+      filename: /var/log/cobald/cobald-tardis.log
+  formatters:
+    test:
+      format: "  %(name)s %(message)s"

roles/cobald/files/cobald-config/tardis.yaml

@@ -0,0 +1,41 @@
+Plugins:
+  SqliteRegistry:
+    db_file: /tmp/drone_registry.db
+  TelegrafMonitoring:
+    host: ed-telegraf
+    port: 8094
+#BatchSystem:
+#  adapter: FakeBatchSystem
+#  allocation: 1.0
+#  utilisation: !PeriodicValue
+#    period: 60
+#    amplitude: 0.15
+#    offset: 0.80
+##    phase: 1.
+#    phase: 1.6
+#  machine_status: Available
+BatchSystem:
+  adapter: Slurm
+  max_age: 0.1
+  options:
+    partition: cobald
+Sites:
+- name: slurmtest
+  adapter: Slurm
+  quota: 20
+slurmtest:
+  # executor: ...
+  StatusUpdate: 0.1
+  MachineTypes:
+    - m1.a
+  MachineTypeConfiguration:
+    m1.a:
+      Walltime: 5
+      Partition: container
+      StartupCommand: /usr/local/bin/start-drone
+      # SubmitOptions: ...
+  MachineMetaData:
+    m1.a:
+      Cores: 3 # cores
+      Memory: 1 # GB
+      Disk: 4 # not passed

roles/cobald/files/slurm-slurmd.def

@@ -4,6 +4,7 @@ From: slurm:slurmd
 %files
 	31-slurmd-configless /etc/docker-init.d/31-slurm-configless
 	/container/volumes/munge/munge.key /etc/munge/munge.key
+	cgroup.conf.noautomount /etc/slurm/cgroup.conf
 
 %post
 	rm /etc/docker-init.d/30-slurmd

roles/cobald/files/start-drone

@@ -4,22 +4,56 @@ export
 echo $@
 nodename=$(hostname | awk '{ print "drone" substr($1,match($1, "([[:digit:]]+)")) }')
 
+SHUTDOWN_DONE=0
+
 function handler_quit(){
+	[ $SHUTDOWN_DONE -ne 0 ] && return
+	set -x
+	echo "drain container"
+	scontrol update NodeName=${nodename} State=DRAIN Reason="cobald node quit"
+	shutdown_jobs=$(squeue -w ${nodename} --noheader -O jobid)
+	[ -n "${shutdown_jobs}" ] && scancel ${shutdown_jobs}
+	#scancel -w ${nodename}
+	i=$(( $(scontrol show config | grep KillWait | \
+			sed 's/^KillWait.*= \([0-9]*\) sec/\1/') - 2 ))
+	while [ -n "$(squeue -w ${nodename} --noheader -O jobid)" -o ${i} -lt 1 ]
+	do
+		i=$(( ${i} - 1 ))
+		sleep 1
+	done
+	scancel -s KILL -w ${nodename}  # hard kill all remaining jobs
 	echo "shutdown container"
+	scontrol update NodeName=${nodename} State=DOWN Reason=shutdown
 	singularity instance stop slurm-drone
 	scontrol update NodeName=${nodename} State=FUTURE
+	umount /inner-cgroup/freezer
+	umount /inner-cgroup
+	SHUTDOWN_DONE=1
 	exit 0
 }
 
-trap handler_quit EXIT
-
 # set -x
 
+trap handler_quit EXIT
+
+echo "mounting cgroups"
+mkdir /inner-cgroup
+mount -t tmpfs none /inner-cgroup
+mkdir /inner-cgroup/freezer/
+mount --bind /sys/fs/cgroup/freezer/slurm/ /inner-cgroup/freezer/
+mount -o remount,ro /inner-cgroup
+
 echo "starting ${nodename}"
-singularity instance start --writable-tmpfs /shared/slurmd.sif slurm-drone \
+scontrol update NodeName=${nodename} State=RESUME # revoke last DRAIN
+scontrol update NodeName=${nodename} State=FUTURE
+singularity instance start \
+	-B /inner-cgroup/:/sys/fs/cgroup/ \
+	--writable-tmpfs /shared/slurmd.sif slurm-drone \
 	slurm-ctl ${nodename}
+# scontrol update NodeName=${nodename} NodeHostname=${SLURM_JOB_ID}
+scontrol update NodeName=${nodename} NodeHostname=${TardisDroneUuid}
 if [ $? -eq 0 ] ; then
-	echo "container started, sleeping $(( 60 * ${SLURM_Walltime}))"
+	echo "container started, sleeping $(( 60 * ${SLURM_Walltime} - 2 ))"
-	sleep $(( 60 * ${SLURM_Walltime} ))
+	sleep $(( 60 * ${SLURM_Walltime} - 2 ))
 fi
 handler_quit

roles/cobald/tasks/main.yml

@@ -10,15 +10,16 @@
     slurm_image_prefix: cobald
     image_name: "{{cobald_image_tag}}"
     dockerfile: "{{ lookup('template', 'cobald.Dockerfile') }}"
-    files:
+    files_list:
-    - dest: cobald-entrypoint.sh
+    - cobald-entrypoint.sh
-      content: "{{ lookup('file', 'cobald-entrypoint.sh') }}"
+    - init-cobaldmodules.sh
-    - dest: init-cobaldmodules.sh
+    - start-drone
-      content: "{{ lookup('file', 'init-cobaldmodules.sh') }}"
+    - 28-sync-container-slurmd
-    - dest: start-drone
+    files: "
-      content: "{{ lookup('file', 'start-drone') }}"
+      {%- set files = [] -%} {%- for i in files_list -%}
-    - dest: 28-sync-container-slurmd
+        {%- set files = files.append(
-      content: "{{ lookup('file', '28-sync-container-slurmd') }}"
+          { 'dest': i, 'content': lookup('file', i) }) -%}
+        {%- endfor %}{{ files }}"
   when: cobald_slurm | default(False)
 
 - name: build generic cobald docker image
@@ -32,13 +33,14 @@
     owner: "{{unpriv_user}}"
     group: docker
 
-- name: copy cobald config (does nothing yet)
+- name: copy cobald config
   copy:
     src: cobald-config/
-    dest: /container/volumes/cobald
+    dest: "~{{unpriv_user}}/cobald/"
+    force: False
     owner: "{{unpriv_user}}"
     group: docker
-  when: False
+    mode: "0644"
 
 - name: ensure network for cobald container exists
   docker_network:
@@ -68,7 +70,7 @@
     repo: https://github.com/thoto/cobald
     dest: "~{{unpriv_user}}/cobald-src"
     version: bugfix/mixed_construction_methods
-    update: no
+    update: no  # FIXME
   become: yes
   become_user: "{{unpriv_user}}"
   register: cobald_git_pull
@@ -78,6 +80,7 @@
     repo: https://github.com/MatterMiners/tardis
     dest: "~{{unpriv_user}}/tardis-src"
     version: master
+    update: no  # FIXME
   become: yes
   become_user: "{{unpriv_user}}"
   register: tardis_git_pull

roles/cobald/tasks/singularity.yml

@@ -23,6 +23,7 @@
   loop:
     - slurm-slurmd.def
     - 31-slurmd-configless
+    - cgroup.conf.noautomount
   register: cobald_copy_sing_files
 
 - name: remove old container

roles/slurm/tasks/host-config.yml

@@ -1,3 +1,6 @@
+# TODO: this does not work quite right since slurm-ctl does not reach the host
+# system. sinfo, scontrol etc. work but srun does not!
+
 - name: "get addresses from docker network"
   docker_network_info:
     name: "{{ docker_network }}"
@@ -28,4 +31,5 @@
 - name: start munge locally
   service:
     name: munge
+    enabled: True
     state: started

roles/slurm/tasks/main.yml

@@ -92,18 +92,11 @@
   notify: reconfigure slurm
   tags: slurm-config
 
-- import_tasks: host-config.yml
-  when: slurm_hostsystem_cluster_access | default(False)
-
 - name: "create docker volume for shared access between nodes"
   docker_volume:
     name: slurm-shared
     state: present
 
-# TODO: reserve some address using docker_network_info and assign as aux
-# address to enable slurmctld to get a static address in order to be
-# reachable from slurm running on docker host to enable submitting jobs.
-
 - name: run slurm docker containers
   docker_container:
     name: "{{ slurm_prefix }}-{{ item.machine }}"
@@ -130,6 +123,10 @@
     label: "{{slurm_prefix}}-{{ item.machine }}"
   tags: slurm-config
 
+- name: configure host system to integrate into slurm cluster
+  import_tasks: host-config.yml
+  when: slurm_hostsystem_cluster_access | default(False)
+
 - name: export facts about slurm cluster to be used by other modules
   set_fact:
     slurm:

roles/slurm/templates/slurm.conf.j2

@@ -105,8 +105,10 @@ Waittime=0
 #MaxMemPerCPU=0
 #SchedulerTimeSlice=30
 SchedulerType=sched/backfill
-SelectType=select/linear
+# SelectType=select/linear
+SelectType=select/cons_res
 #SelectTypeParameters=
+SelectTypeParameters=CR_CORE
 #
 #
 # JOB PRIORITY